Closed janhurst closed 4 years ago
I was thinking on a few variables as below,
CTDone | Any head CT performed? | YesNo 0 No 1 Yes | This indicates if any head CT was performed regardless of where it was obtained except if the ED head CT was marked as "not interpretable". |
---|
PosIntFinal | Clinically-important traumatic brain injury | YesNo 0 No 1 Yes | Clinically-important TBI was defined as having at least one of the following: (1) neurosurgical procedure performed, (2) intubated > 24 hours for head trauma, (3) death due to TBI or in the ED, (4) hospitalized for >= 2 nights due to head injury and having a TBI on CT. |
---|
CTDone Any head CT performed? YesNo 0 No 1 Yes This indicates if any head CT was performed regardless of where it was obtained except if the ED head CT was marked as "not interpretable".
I don't think CTDone is a class variable. It is only recording if a CT was actually taken. I also doon't think it would be an input into the model, as we won't know this information at the time of assessing a new patient.
We could consider if we should only train the model for records where a CT was actually taken, but we need to then be careful if there are any records where TBI was found but no CT was conducted.
The final variable in the dataset is PosIntFinal, which indicates if a clinically important TBI was found regardless of a CT scan.
PosIntFinal is a composite calculated from DeathTBI, Neurosurgery, Intub24Head, and HospHeadPosCT.
These five variables are effectively the candidate class variables.
I think we should try to predict PosIntFinal before we worry about trying to have a more sophisticated model that predicts which outcome determined PosIntFinal. I'm going to update my TPOT PR with this and I'll link it here when its ready.
PosIntFinal was decided as the Target variable.
I'm working on this and provide more details this weekend.